An Evolutionary Multitasking Method for High-Dimensional Receiver Operating Characteristic Convex Hull Maximization

Maximizing receiver operating characteristic convex hull (ROCCH) is a hot research topic of binary classification, since it can obtain good classifiers under either balanced or imbalanced situation. Recently, evolutionary algorithms (EAs) especially multi-objective evolutionary algorithms (MOEAs) have shown their competitiveness in addressing the problem of ROCCH maximization. Thus, a series of MOEAs with promising performance have been proposed to tackle it. However, designing a MOEA for high-dimensional ROOCH maximization is much more challenging due to the “curse of dimension”. To this end, in this paper, an evolutionary multitasking approach (termed as EMT-ROCCH) is proposed, where a low-dimensional ROCCH maximization task
$T_{a}$
is constructed to assist the original high-dimensional task
$T_{o}$
. Specifically, in EMT-ROCCH, a low-dimensional assisted task
$T_{a}$
is firstly created. Then, two populations,
$P_{a}$
and
$P_{o}$
, are used to evolve tasks
$T_{a}$
and
$T_{o}$
, respectively. During the evolution, a knowledge transfer from
$P_{a}$
to
$P_{o}$
is designed to transfer the useful knowledge from
$P_{a}$
to accelerate the convergence of
$P_{o}$
. Moreover, a knowledge transfer from
$P_{o}$
to
$P_{a}$
is developed to utilize the useful knowledge in
$P_{o}$
to repair the individuals in
$P_{a}$
, aiming to avoid
$P_{a}$
being trapped into the local optima. Experiment results on 12 high-dimensional datasets have shown that compared with the state-of-the-arts, the proposed EMT-ROCCH could achieve ROCCH with higher quality.